물리 범위·스키마 보장· Bounds-safe & schema-conform
100%
75%
by construction
의도 일치도· Correct vs. user intent
91%
75%
fuzz inputs incl. out-of-range, multi-number, malformed
출력 페이로드· Output payload
≈53 B
25–86 B
fixed normalized vs. variable
파싱·컴파일 오버헤드· Parse / compile overhead
≈7 µs
—
LLM call cost is identical on both paths